Bumjun Kim

peer-reviewed publications

(*) denotes equal contribution, (†) denotes corresponding author.

1. ICML

   DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs

   Bumjun Kim^*, Dongjae Jeon^*, Moongyu Jeon^*, and Albert No^†

   In the International Conference on Machine Learning, 2026.

   PDF

   Parallel decoding for diffusion LLMs using attention-derived dependency structure.

2. ICLR

   Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs

   Bumjun Kim^*, Dongjae Jeon^*, Dueun Kim^*, Wonje Jeung, and Albert No^†

   In the International Conference on Learning Representations, 2026.

   OpenReview

   A method for reducing early termination in instruction-tuned diffusion language models.

3. CVPRFindings

   Memorization In Stable Diffusion Is Unexpectedly Driven by CLIP Embeddings

   Bumjun Kim and Albert No^†

   In the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026 Findings.

   PDF

   An analysis of Stable Diffusion memorization that identifies CLIP embeddings as a key driver.

4. NeurIPS

   A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse

   Moongyu Jeon^*, Sangwoo Shin^*, Bumjun Kim, Kyelim Lee, and Albert No^†

   In Advances in Neural Information Processing Systems, 2026.

   OpenReview

   Identifies any-order masked supervision, full attention, and position-invariant value vectors as key mechanisms behind reversal-curse mitigation.

5. EMNLP

   Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

   Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Yoonjun Cho, Dongjae Jeon, Bumjun Kim, Jean Oh^†, Youngjae Yu^†, and Albert No^†

   In the 2026 Conference on Empirical Methods in Natural Language Processing (Main Conference).

   PDF

   A benchmark revealing the fragility of vision-language reward models to paraphrases of identical robot trajectories.

preprints

1. arXiv

   Low-Confidence Remasking Traps Flexibility: Realizing Arbitrary-Order Potential for Diverse Rollouts in Diffusion LLMs

   Moongyu Jeon^*, Dongjae Jeon^*, Bumjun Kim, Mingyu Kim^†, and Albert No^†

   arXiv preprint, 2026.

   PDF

   Shows that low-confidence remasking, rather than arbitrary-order generation, suppresses rollout diversity, and introduces entropy-guided initialization to improve exploration.

workshop papers

1. ICML-W

   Single-Step Initialization for Exploratory Parallel Rollouts in Diffusion LLMs

   Dongjae Jeon^*, Bumjun Kim^*, Mingyu Kim^†, and Albert No^†

   In the ICML Workshop on Structured Probabilistic Inference & Generative Modeling, 2026.

   OpenReview

   Improves exploration in dLLM RL post-training by unmasking one randomly selected position before decoding to diversify parallel rollouts.